[hibernate-dev] What should happen to queued operations when a collection is detached?

[Gail Badner]

Since I haven't heard anything back, I've gone ahead and updated my PR [1]
to ignore queued operations when merging a detached collection. This restores
the functionality that was in place prior to HHH-5855 (which caused
HHH-11209).

In addition I've added warnings for the following situations:
* a detached collection is being merged that has queued operations;
* a collection with queued operations is attached to a session;
* a collection with queued operations is detached from a session.

I've also added a test case that shows that Session#merge and
Session#saveOrUpdate have different results when called on an entity with a
collection with queued operations. When Session#merge is called, the queued
operations are ignored. When Session#saveOrUpdate is called, the queued
operations are executed on flush.

I don't like this inconsistency, but I'm not sure what to do about it. I
don't think it's a good idea to have collections floating around with
queued operations. As far as the application knows, it is just an
uninitialized collection. I can see that the entity could find its way into
a new session, with surprising results.

IIUC, an application probably shouldn't detach an entity with a collection
in this state (having queued operations), so I think it makes sense that a
warning be logged when this happens. On the other hand, the warning will
also be logged if a transaction involving updates to a collection with
queued operations fails.

I would really appreciate some feedback on this, since HHH-11209 is a bad
regression that really should be fixed.

Thanks,
Gail

[1] https://github.com/hibernate/hibernate-orm/pull/2460

On Fri, Sep 28, 2018 at 6:27 PM Gail Badner <gbadner at redhat.com> wrote:

> HHH-11209 involves a bug merging a detached entity with an uninitialized
> collection that has queued operations.
>
> After looking into this issue I found a bug
> where AbstractPersistentCollection#operationQueue was not being cleared
> after a commit, if there was no cascade mapped for the collection.
>
> I've fixed that in a PR. [1]
>
> After that fix, the detached entity in the test case attached to HHH-11209
> has an uninitialized entity without any queued operations after commit,
> which can be merged successfully in a new session/transaction.
>
> There is still a problem when Hibernate tries to merge a detached entity
> has an uninitialized collection with queued operations though. My PR throws
> UnsupportedOperationException in this case, and there is a test case that
> reproduces it.
>
> I've been working on a fix to add support for this, but I have my doubts
> that it really should be supported. I see that prior to fixing HHH-5855
> (which caused HHH-11209), Hibernate simply ignored the queued operations in
> the detached collection. [2].
>
> The fix for HHH-5855 properly dealt with merging a managed collection that
> was uninitialized with queued operations. It introduced the bug where a
> NullPointerException would get thrown if that collection  was detached.
>
> If we want to support merging the queued operations, then I would consider
> that an improvement. Would this be a worthwhile improvement? I'm guessing
> not, but I wanted to get some opinions.
>
> For now, I see a couple of ways to deal with this so that HHH-11209 can be
> wrapped up:
> 1) ignore queued operations in a detached collection when merging, as was
> done before HHH-5855 was fixed.
> 2) clear the queued operations when the collection is detached.
>
> Comments?
>
> Thanks,
> Gail
>
> [1] https://github.com/hibernate/hibernate-orm/pull/2460
> [2]
> https://github.com/hibernate/hibernate-orm/commit/c1934b72edb4f781520937618b3b750bebb84576#diff-2c7912a47063b9ea81323bf9c6628895L651
>